Prediction of next price move doesn't mean positive pnl... in any case interesting research

Interesting paper, thank you for bring it to the group's attention. I personally find it very worthwhile to read academic papers like this one; they make you think and may in fact spark an idea that improves your trading.

P.S.

As I was a student I learnt a lot from the book "Financial Modelling with Jump Processes" by Cont and Tankov

Dear Prof. Cont, the topic is pretty interesting, but:

1) According to the manifest of reproducible research,

https://statweb.stanford.edu/~wavelab/Wavelab_850/wavelab.pdf

"An article about computational science in a scientific publication is not the scholarship itself, it is merely advertising of the scholarship. The actual scholarship is the complete software development environment and the complete set of instructions which generated the figures".

In your case it may be a problem to conform to this manifest due to legal issues (data redistribution) but anyway...

2) You claim that the LSTM-model outperforms the linear (VAR) model about 10%. But why do you "compare the neural network against a linear model for approximately 500(!) stocks", whereas "Using a Deep Learning approach applied to a large dataset of billions of orders and transactions

for 1000(!) US stocks"?

3)If the results are really so promising as you affirm, it is unclear why do you publish it instead of trying to sell it to proprietary

trading companies of hedgefonds?

Probably because you results are like the glorified paper by "Foundations of Technical Analysis: Computational Algorithms, Statistical Inference, and Empirical Implementation" by Lo, Mamaysky and Wang?

(interesting but hardly applicable in practice)

https://letyourmoneygrow.com/2018/03/11/patterns-of-technical-analysis-do-they-work/

Or is it merely an attempt to exploit the buzz-words?

https://letyourmoneygrow.com/2016/10/23/big-data-and-deep-learning-a-technology-revolution-in-trading-or-yet-another-hype/

The idea is good, practical aspects of it in production needs to be seen.

Thanks for sharing. Definitely an interesting piece of research - although I was secretly hoping to see the CNNs applied to the spatial structure of the order books.

it would be very interesting to see to which extent the universality feature ( universal model beating the stock-specific one) survives the increase of the horizon of the forecast. It might require another supercomputer grant to find out...

I suspect the data was provided only for academic purposes (it is mentioned that the supercomputer access is part of a CFM grant), but for professional investors there is a UK company that provides a similar set up (I have no affiliation with them): https://bmlltech.com/#our-product they 'rent' the whole limit order data + cloud computing, and looks remarkably similar to the architecture described in the paper.

https://www.mql5.com/en/users/ucakci

Do you want to earn money ? Check my mql profile

As Mattia mentioned, prediction of the next price does mean positive PNL (if you start from scratch you need first to buy/sell the stock at t+1 and take profits at t+2 - so you need to forecast two steps ahead), but it might be useful for flow traders who have a block execution to do. The interesting bit is that the trader can model into the NN the order he will introduce (and which will modify the forecast - some sort of observer effect - he can see the effect of his order into the price)

Great application of deep learning.